31 lines
1.3 KiB
Markdown
31 lines
1.3 KiB
Markdown
# Ollama gpt-oss-20b Options Training (RTX 5060 Ti)
|
|
|
|
This repo contains the full workflow, scripts, and runbook to fine-tune
|
|
`openai/gpt-oss-20b` on options-trading ebooks and deploy the merged model into
|
|
Ollama as `trained-options-model` on a TrueNAS SCALE box with an NVIDIA GPU.
|
|
|
|
Start here: `AGENTS.md` (detailed step-by-step, failures, fixes, commands, and
|
|
paths).
|
|
|
|
## Contents
|
|
- `tools/` scripts for extraction, relevance filtering, dataset building, and
|
|
LoRA fine-tuning.
|
|
- `training_data/` curated dataset + LoRA outputs used for training.
|
|
- `remote/ollama/` copies of the remote Modelfile and template used to enable
|
|
tool calls in Ollama.
|
|
- `Modelfile.trained-options-model` (local reference; see AGENTS for the exact
|
|
remote Modelfile used).
|
|
|
|
## What is NOT included
|
|
- Raw ebooks (`eBooks/`) are intentionally excluded.
|
|
- Base model weights are not committed; download via Ollama or Hugging Face as
|
|
described in `AGENTS.md`.
|
|
|
|
## Quick pointers
|
|
- Training runs on TrueNAS SCALE (GPU via TrueNAS Apps/middlewared).
|
|
- Ollama runs in a TrueNAS App and stores models under
|
|
`/mnt/fast.storage.rushg.me/datasets/apps/ollama.models`.
|
|
- Tool-call support requires the correct TEMPLATE block (see `remote/ollama`).
|
|
|
|
For full instructions and exact commands, see `AGENTS.md`.
|